UC San Diego

The quest to understand and reverse engineer natural intelligence has captivated humanity for centuries. While we have made breakthroughs in neuroscience and machine learning individually, the intersection of the two remains largely unexplored and this dissertation aims to tap into its unknown potential. Using a dual modeling-experimental methodology, I study the cognitive capabilities of the prefrontal cortex, a brain-region essential for higher-order cognitive functions.I first define the postulates of lifelong learning, the idea of a continually learning algorithm capable of mastering innumerable tasks sequentially with efficient transfer of knowledge (Chapter 1). I discuss the state-of-the-art solutions that have been proposed in engineering and their shortcomings. I subsequently describe the systematic approach that I undertake over the course of my PhD. I first look at the role of the prefrontal cortex in the modulation of behavior across social contexts. I use state space analysis to identify distinct neural representations within the PFC linked to social rank and competition outcomes. I then devise a novel adaptation of the Hidden Markov Model to identify latent states associated with flexible social behavior, revealing potential mechanisms for context-specific neural activation within the PFC. Function and disease are two sides of the same coin. In order to broaden my understanding of prefrontal mechanisms, I next pursue a collaborative study focused on the detrimental impact of chronic stress on PFC function and behavior (Chapter 3). I identify a method to categorize animals based on stress-induced anhedonia. I use state space analysis to identify distinctions in the anhedonic state from the hedonic in PFC activity and study its evolution over a chronic stress paradigm. Given the findings from previous chapters, I develop a hierarchical modular neural network architecture based on the PFC, integrating internal state dynamics and a novel neuromodulatory filtering mechanism (Chapter 4). This model successfully reproduces animal behavior and PFC function while being simulated on a contextual foraging task. This underscores its validity and potential for understanding the neural underpinnings of lifelong learning as well as related neuropsychiatric conditions. Through this dissertation, I aim to advance our understanding of the neural basis of lifelong learning. The novel computational models and experimental findings presented here have far-reaching implications for the development of artificial intelligence systems with enhanced cognitive capabilities, as well as for potential therapeutic interventions targeting PFC dysfunction in various neuropsychiatric conditions.